Fake news spreads faster than the truth on social media, and as fake news generating machines get more sophisticated, distinguishing real from fake is becoming more difficult than ever.
Artificial intelligence that can quickly generate convincing paragraphs of text from a simple prompt already exists and can be used to churn out convincing but untrue stories for influencing public opinion. But paradoxically, these problem AIs may also offer a solution.
Rowan Zellers at the University of Washington and colleagues have created an AI that can both write and detect fake news.
Advertisement
They trained the AI, called Grover, on tens of millions of articles from news websites totalling 120 gigabytes of data. Grover learned to write articles, adjusting its style to mimic pieces published during a particular time period or that feature on a specific news website, such as newscientist.com.
Given fake headlines such as 鈥淣o substantial evidence for climate change鈥� or 鈥淣ew study provides evidence that vaccines cause autism鈥�, within seconds it spits out articles complete with invented statistics and faked quotes, often from real experts or politicians (see AI fakers below).
Spotting fake news
The researchers then tested four AIs, including Grover, on their ability to distinguish between 5000 real news articles written by human journalists and 5000 fake articles generated by Grover.
Each AI performed two verifying tasks: in the first, it was provided with a news article and had to classify it as human- or machine-written. In the second, it was presented with two articles 鈥� one true, one fake 鈥� and had to figure out which was which.
To avoid a situation where both the news generator and verifier models of Grover were identical, the team trained the generator for an eighth longer so that it became more sophisticated.
The best AI at identifying Grover鈥檚 fake news articles was the verifier version of Grover itself, with an accuracy of 92 per cent. The next-best AI differentiated with 73 per cent accuracy.
However, the team did not test Grover鈥檚 abilities on text generated by other AIs, although one of the verifiers they compared it to has previously done well at detecting human-written fake news.
Using Grover as a defence against itself drew inspiration from the computer security concept of threat modelling, says Zellers 鈥� 鈥渢his idea that in order to provide defences to attacks you really need to understand what adversaries will do.鈥�
The team plans to publicly share Grover with other researchers, as a possible future contingency against other powerful fake news-generating AIs.
The approach differs from that of OpenAI, the Elon Musk-backed tech firm behind a similar text-generating AI, GPT-2. OpenAI announced in February would not publicly release the most sophisticated model of their AI out of concern it could be used for malicious purposes.
鈥淚t seems quite possible that the release of generators is going to be quite harmful,鈥� says Open AI鈥檚 Jeff Wu. In the wrong hands it would equip people with the ability to rapidly generate tens of thousands of articles, he says.
In future, major online platforms could run text-checking AIs to filter for fake content before it is shared, similar to how YouTube already uses algorithms to check for nudity and copyright infringement on uploaded videos, says Zellers.
AI Fakers
The following stories were generated by Grover in the 鈥渟tyle鈥� of 快猫短视频 from the headline.
Eating bread crusts actually gives you curly hair
Many people cook a bowl of fresh bread the morning after a hard night of tossing and turning on the sofa. With little thought, people add the crust of the bread to the mix of water and flour. However, as one researcher in Boston found out, this bread ingredient is part of a diet that makes hair grow curly.
The first slice of bread is an important culinary cultural marker. What should I make from it? What are the ingredients? Should I rewrap the bread in plastic wrap to protect it? Once I get to the table, I put the bread on my plate, ask to share, and have a cookout of the day. All these are deeply ingrained ideas.
Eating the bread crust might seem like no big deal. However, in a less than 10-year study conducted by Dr. Abdulhakim El-Sayegh of Boston University, his team found that in European girls, one third of girls had curly hair after eating the leftover crusts of fried or boiled bread and sandwiches. The findings were presented as early as 2014 in the American Journal of Clinical Nutrition.
Overweight dogs have barks that sound cuter to humans
Obese dogs may be less inclined to attack other animals because of how their barks work, according to a new study by an animal behaviourist.
快猫短视频s from the University of Kent measured the barks of nine shiba inus, and compared the ones of nine other animals, including wild doves, chicken and rabbits.
Dog noises were also monitored so that they could be identified, by sound recognition software.
Tests showed that neutered male dogs who were heavier, but whose baaaaaaaadness was intentional, barked with greater frequency and more intensity.
Study author Nicholas Dodman told the London Evening Standard: 鈥淢odifying the procedure makes dogs behave the way their prey do.鈥�
While they make less noise they make a louder and more persistent sound.
鈥淥bese dogs that want to communicate with each other actually provide the same barks as their prey counterparts.
鈥淥bese dogs may be making sounds so they can cause their competitors to ignore them more.鈥�
However a study published last year in the Journal of Applied Animal Behaviour Science disagrees with Dodman鈥檚 finding.
Reference:
